The Semantic Web in an SMS

Onno Valkering, Victor de Boer, Awa Gossa Lô, Romy Valkering, Stefan SchlobachVrije Universiteit Amsterdam

Can the (Semantic) Web (be made to) mean something for knowledge sharing even under very constraining conditions?

No internet, no computer, no electricity

Multitude of languages, levels of literacy

Web for Development challenge

http://worldwidesemanticweb.org

Low-resource knowledge sharing platform

Low-power, ubiquitous, cheap hardware FLOSS components

Rapid prototyping + deployment of (knowledge-intensive) services

User InterfacesVoice servicesSMS-based Visual

Wifi, 3g or GSM network

RDF Data store (Linked Data) allows for flexible data integration across applications, deployments

allows for easy development of information services relevant to local communities in their preferred language

Kasadaka (“talking box”)

Data store

GSM/Voice interfaceWeb Interface

Text-To-Speech / ASR

Radio users

Field operative

Users

Information Services

Market Information System (Mali)

Veterinary service (N-Ghana)Knowledge base and diagnosis system(CommonKADS)Information sharing across locations

Poultry vaccination service (Mali)

Seed market (Mali and Burkina Faso)Seed quality

Current cases

Machine to machine information integration

GSM network

SPARQL in an SMSConverters to translate SPARQL HTTP request to SMS message (140 or 160 chars) and vice versa

CONSTRUCT, INSERT/DELETE DATA

ChallengesBlending synchronous and asynchronous messagingSPARQL/ RDF compressionUnpredictable query result sizes

Compression for small datasets experiments

StrategiesDifferent serializations: RDF/XML, N-triples, Turtle, HDT1, EXI2 Compression (zip)Assume shared vocabularies (top 20 from prefix.cc) and remove redundant (inferenced) triples (RDFS reasoning)

Evaluated on real-world datasets LOD Laundromat3

232,822 small datasets (1-1,000 triples)

[1] Fernández et al. “Compact Representation of Large RDF Data Sets for publishing and exchange” (ISWC 2010)[2] Käbisch et al. “Standardized and Efficient RDF Encoding for Constrained Embedded Networks” (ESWC2015)[3] http://lodlaundromat.org/ and Rietveld et al. “LOD Lab: Experiments at LOD Scale “ (ISWC2015)

Compression experiments resultsNumber of SMSes

Avg. number of triples

1 0

2 3

3 8

4 16

5 24

6 84

7 98

8 126

9 189

10 301

For very small datasets (<40 triples), n-triples + gzip works best For larger datasets Turtle+gzip compresses bestRemoving redundancies using shared vocabularies adds additional compression

1-10 11-20

21-30

31-40

41-50

51-60

61-70

71-80

81-90

91-100

101-200

201-300

901-1000

0

10

20

30

40

50

60

70

80

N-triples +gzipTurtle+GzipBest + vocabulary-based

Size of dataset in triples (binned)

Com

pres

sion

(per

cent

age

wrt

n-tr

iple

s)

Evaluation: 4 scenarios in 2 cases Digivet and RadioMarche applicationsFour scenarios / SPARQL queries in total

Results

Conclusions

Semantic Web over SMS is feasible using data compression + semantic background knowledgeeconomically feasible for ICT4D services for small datasets

Semantic Web without the Web is possible cf. IOT

Knowledge engineering, knowledge sharing for allTowards a Computer Science for Development (CS4D)

Thank you

github.com/onnovalkering/sparql-over-smsgithub.com/abaart/KasaDaka

kasadaka.comworldwidesemanticweb.org

w4ra.org

v.de.boer@vu.nl

TriplesN-triples +gzip

RDF/ XML

RDF/XML+Gzip Turtle

Turtle+G zip HDT

HDT+ Gzip EXI

EXI+ Gzip

Best + vocabulary-based

1-10 50.7 103.8 77 102 70.3 495.5 180.1 57.5 65.9 44.211-20 22.5 62 27.1 50.5 24.2 122.2 47 23.3 24.9 18.921-30 16.2 58.2 18.5 48.7 16.3 79.5 31.1 16.5 17.5 13.631-40 28.3 69.1 30.9 62.1 28.6 86.5 40.7 28.2 29.1 23.541-50 9.8 51.2 10.2 42.3 8.6 38.1 14.8 9.3 9.7 851-60 17.2 59.2 17.5 50.1 15.9 50.5 22.8 15.8 16.3 8.761-70 11.8 58.5 12.4 42.4 10 43 17.7 11.1 11.6 671-80 8.8 54.8 8.5 40.9 7 31.6 11.2 7.5 7.8 6.481-90 6.7 52 6.3 40.6 5.1 25.4 9.1 5.8 6 4.491-100 8.1 54.9 7.6 40.4 6.2 26.9 9.7 6.8 7 5.7101-200 8.8 62 8.3 39.2 6.7 24.7 10.1 7.6 7.9 5.7201-300 4.8 50.8 3.6 39 2.8 13.4 4 3.6 3.6 2.7301-400 4.8 51.5 3.3 37.7 2.5 11.4 3.3 3 3.1 2.5401-500 4.4 51.5 2.9 37.4 2.2 10.4 2.7 2.6 2.7 2.2501-600 5 53.8 3.4 38.7 2.5 8.9 3 2.9 3 2.4601-700 4.1 51 2.5 35.9 1.7 8.5 2.2 2.3 2.4 1.7701-800 4.5 51.1 2.7 36.2 1.9 8.1 2.1 2.4 2.4 1.9801-900 4.4 51.1 2.6 36.4 1.8 7.9 1.9 2.3 2.3 1.8901-1000 4.1 50.9 2.4 36.5 1.7 7.7 1.7 2.1 2.1 1.7801-900 4.4 51.1 2.6 36.4 1.8 7.9 1.9 2.3 2.3 1.8901-1000 4.1 50.9 2.4 36.5 1.7 7.7 1.7 2.1 2.1 1.7

For very small datasets (<40 triples), n-triples + gzip works best For larger datasets Turtle+gzip compresses bestRemoving redundancies using shared vocabularies adds additional compression

SPARQL in an SMS

Enable (Semantic) Web data exchange over GSM networks.

Practical differences HTTP and SMS:SMS works with phone number, HTTP works with URLs.SMS has a size restriction, HTTP practically has none.SMS is one-way messaging, HTTP follows request-response.

Basic M2M communication based on SPARQL.